https://www.bridgebench.ai/nerf-bench
They test it on launch day, then benchmark it against that. A deviation of above 10% is considered a change. They're currently tracking Opus 5.5 and GPT-6 Astra.
This bench famously detected a degradation of Opus 4.6 which Anthropic later blogged about. I personally think people sense nerfs more often than they happen and that it's often about honeymoon effects.
It doesn’t mean the big labs don’t also nerf models! But if they didn’t you’d still have users complaining.
`percieved_performance = actual_perf/expectation`
`expectation` is an increasing function over time.
`actual_perf` is a stochastic function of the model's true ability, context, etc. -> a recipe for some bad sessions.
As for multiple bad sessions in a row, this is a studied phenomenon in gambling where players perceive "runs" because our brains love to find patterns.
e.g. The original apple shuffle and the Risk app ins which a string of songs from the same album or three one roles are "not random"
How to Shuffle Songs? - https://web.archive.org/web/20220215030739/https://engineeri... ( https://news.ycombinator.com/item?id=38330877 78 points, 65 comments)
Took a little bit of digging to find it - I remembered the graphic at the top and found a blog post that copied it and linked to the blog post, but the blog post isn't there anymore... so web archive.
The current version of the blog post is from 2025 - https://engineering.atspotify.com/2025/11/shuffle-making-ran... (which didn't get any traction on HN)
Also, consider that in flipping 10 coins, you'll find strings of 2 heads in ~86 percent of runs, 3 in ~51% of runs, 4 in ~25% of runs, 5 in ~11% of runs...and in strings of 100 flips you'll finds strings of 6 in ~55%, 7 in ~32%, 8 in ~17%, 9 in ~9%...
Widening the range from "rolling exactly 15" to "rolls 15 or 16" or "rolls between 14-17" makes the strings even more likely as you're doubling the success rate from "only 9 15s" to the "any string between 9 fifteens, through 16 and 8 fifteens, to 9 16s" space.
To check if your random is randoming you can calculate expectations versus your results (using a large enough sample) with:
For N samples of a fair die, expexted runs k with probability of success p and failure q can be calculated as:
General Variables: N = total number of rolls/trials k = target streak length p = probability of getting the target outcome (e.g., 1/20 for a specific roll on d20 or 1/10 for two specific results) q = probability of getting any other outcome (1 - p)
Expected runs of AT LEAST length k: E(runs >= k) = p^k * (1 + (N - k) * q)
Expected runs of EXACT length k: E(exact k) = p^k * q * (2 + (N - k - 1) * q)
Personally, I find that 'sticky' dice always provide a nice narrative device, at least in narrative games. A character who's player can't seem to roll over a 10 must, after all, be cursed or perhaps deliberately sabotaging the party.
Been through that this week as well with 100% success on 40% odds over multiple iterations on my game.. I tend to not dig into random but just rather ensure it works 'as closely to intended' as possible..
I was like, wow, I guess the shoes must be literal torture devices full of MRSA-covered broken glass at this point. They've been getting continuously worse for 24 consecutive years!
Of course, what was really happening is that they were not getting worse, but naturally every year there was some small percentage of vocal dissatisfied users, while the silent majority simply enjoyed their shoes and didn't have much to say about them.
(The sorta-opposite happens in sneaker reviews as well. People will gush about how cushy the sole in some particular new sneaker is. Well, yeah, of course it's cushy -- you're comparing a new sneaker to your old sneaker where the foam had lost its bounce...)
The first real issue where i wanted to leave was the Claudish nonsense. If not for 5.5 i'd be on OpenAI by now.
Ikea is notorious for this: The early Billy bookcase had heavier veneer and sturdier construction early on, and was actually a really good purchase for the money. The later years replaced veneer with paper foil, used thinner shelves, frames, and backing panels, and was just significantly weaker.
I (used to) buy pre-spliced/terminated fiber optic cables with some frequency from Amazon and came to be familiar with the brands and their quality. One time while shopping for some fiber optics, I saw Amazon Basics-labeled OM-3/OM-4 MMF cable at a very tempting price, so I purchased some to see if it was any good.
To my utter shock and surprise, when I received the trademark plain cardboard boxes with the Amazon Basics label on them and proceeded to open them, I found that I was sent boxes of cables still factory wrapped with labels that clearly read “Corning Optical” – which if you know anything about optical fiber, was pretty much the premium brand in the game. I should have stocked up because the next time I went to order I found out their experiment had ended and they no longer sold “Amazon Basics” finer cables.
Amazon Basics AA NiMH was well known to test exactly the same as the top Japanese brand 'Eneloop'. Extremely good specs all around
Recently though, they are still called Amazon Basics but no longer test like Eneloop. They've changed manufacturers for the worse and are hoping no one notices...
In the interests of saving my sanity and time (it's not free!) having to chase down which batch of which brand is "good" at the moment, I just decided on Eneloop all the time. Sure, we now have like 100-120 or something (wife likes flameless candles - just bought another 16-pack AAs) and I COULD maybe have saved $200 by buying dirt cheap. But all the time spent debugging flaky batteries, having the spouse complain, etc wasn't worth it to me (I get paid reasonably well).
I actually made a battery tester as a hobby electronics project. Fully dumps the energy while measuring mAh and seeing some measurements of internal resistance.
Something I did notice was that crappy battery chargers can permanently damage even Eneloops. So don't cheap out on the chargers either.
The high speed chargers (2 hours or less) are on the edge of what is safe and could permanently damage a cell. The 4 hours or slower chargers are just way more reliable. There's probably someone out there testing different chargers (trying to find the safe fast chargers) but given how cheap NiHMs are (even the "expensive" Eneloops), it's just easier to use 4 hour or 8 hour chargers and have masses and masses of extra NiHMs laying around.
My example is like if you opened the Amazon Basics box and found Eneloop branded white/black/blue batteries directly.
Shrinkflation is a thing, which people suddenly started noticing in the past 5 years.
There's also "the Schlitz Mistake", which I've heard summarized as "most customers won't notice if you take your product's quality from A to B (or C), but they definitely will if you take it from A to J (or A to M)"
At this point I kinda assume that any company releasing year updates to a physical product that _doesn't_ take the opportunity to trim costs / reduce quality would be vulnerable to a shareholder lawsuit for leaving money on the table...
and
On the other hand that idea that "companies exist to make money for shareholders, to ONLY make money for shareholders, and doing any other than maximizing shareholder returns is bad" is pretty commonly accepted (and, I believe, enshrined in US law)
It's been much longer than that: https://en.wikipedia.org/wiki/Toblerone#2016_size_changes
Summarizes the video game industry pretty well
most running shoe series get heavier year to year as manufacturers turn to cheaper materials and add cushioning to try to attract more adopters
it's almost universal, very few manufacturers seem to be able to resist tampering
(heavier shoes are slower, every three ounces is equal to another vo2max point lost)
The foams are getting better, the shoes lighter, they are more cushioned and more responsive in general. Especially the ASICS.
you mean NEW MODELS are being introduced with lighter faster foams
not the same model year to year
modern example: Saucony Endorphin Speed
v1 in 2020 was award winning
v2 in 2021 was almost the same, more praise
v3 bleh
v4 v5 bleh bleh
they cannot resist tampering
Wow the old grumpiness that lingers in my head from losing my favorite shoe. Who knew? Now I'm old and heavier and run in the Nimbus and am happy again. But you're right, of course, that most of the model changes are just fine and people like to complain.
This is true of all reliability and performance paradigms, incidentally
I believe in temporary nerfs. Operators reducing quality significantly to increase throughput for whatever reason (high demand?). Same session, model being completly incapable, and it being back to normal the next day. Experienced that with Anthropic models way too many times. Never on weekends, usually during US work days.
It's been a few months since I last recall this though, must have been pre-opus-5. And I know there are benchmarks for this as well, hence "believe".
People called it lazy mode, in that instead of writing the script you asked for it would basically tell you to learn to code then check out xyz topics to tackle the problem.
You have to pretend to be angry in such situations?
I tell the LLM what to do in its loop. It builds it. I tweak it until it is perfect. I care a lot about UI.
I'm mostly building my own UIs lately, but I find no problem with the development loop. I'm building much better UIs, because it's way easier to test things, and scrap things that I thought would work, but don't. It all happens in a matter of minutes.
Humans should use more llms.
Anthropic cut a deal with SpaceXAI in May - $1.25b/mo. Before that, they employed months of dishonest nerfy strategies, to an extreme.
This is what pissed me off the most. Make it slower, rate limit it, move the credits to other time slots, idk.. but returning BAD results? That's the worst approach you could take.
This alone makes the benchmark unsound.
Deepseek 4.1 ranks very low in this benchmark but it has proven so capable that after being simultaneously on Max x20 and Pro x20 subscriptions, i've transitioned to using DS 4.1 as a daily driver and am very satisfied.
My point is, i think their overall ranking makes sense, matches my experience with out of the box capabilities for vague and underspecified tasks. But seeing a model rank low in their ranking does not mean that the model is incapable. Having skills and guidelines has a lot of influence on what you get out of a model.
I'd love to hear more, I'm considering jumping ship. I'm running a Debian desktop if that's a concern.
Except maybe that the model often believes that he is running out of context, and needs to rush so i occasionally need to jump in to tell it that it still has plenty of room left.
But this does not degrade the quality of my overall experience in a meaningful way.
It's meant for their test env I think, so is not documented to my knowledge
https://code.claude.com/docs/en/monitoring-usage#usage-monit...
thanks for correcting me on that regard
I can't imagine sticking with a supplier that plays games like that with me. Tokens are a pretty vague quantity to begin with (you don't control how many tokens a model puts out in response) and giving a couple of purposefully wrong responses will happily inflate your bill, but you don't care because eventually it worked. It's almost an ideal vehicle to scam people.
Imagine the power company being able to decide how much you consume and at which price point.
Airlines, banks, health insurance…
> Step 2: Regulatory Capture
is being worked towards.
Are they though? Or is it just what some companies would want them to be?
I don't know if that's the actual origin of the term nerf, but it was the first time I'd heard it.
When you're running something at a loss, you can mistreat your customers and they'll still stick around (I'm an example). OpenAI and Anthropic are now cheaper than Chinese models on subscriptions, while being 6-10x more expensive on the API.
My guess is they need the user numbers for the IPO and are willing to take a temporary loss in the meantime. By the time they go public, they'll either drop the subscription model or it'll turn into what the Chinese providers already offer: basically just a cap on how much API you can consume. Same same.
It's not clear what API tokens actually cost them, but I looked into running a local model, and it's way outside the budget of an individual or even a small or medium business (hundreds of thousands of dollars). So my guess is that running these models economically isn't possible, even if they're delivering real business value (coding, research, etc.). In other words, at API prices I'd just stop using AI, and I suspect most other developers would too.
Datacenters have massive economies of scale. Everything from cheaper electricity to having specialized, more efficient hardware to simply being able to run it continuously at near-100% utilization, all adds up.
Many things in the economy - most notably, manufacturing of most consumer goods - only makes economic sense once you're producing for/serving millions of people. This is not unusual.
> In other words, at API prices I'd just stop using AI, and I suspect most other developers would too.
Many say that, but I sincerely doubt they'd actually follow through. People might get more conservative about how they spend their tokens, but AI today is just too good at eliminating drudgery and boring / bullshit parts of daily work to give up on merely 3-5x price increase.
Sure. Issue is, no one is providing on how much it actually costs to burn these tokens. And as we don't know, we can only speculate.
> Many say that, but I sincerely doubt they'd actually follow through.
I have a $100 open ai sub and I track my token usage. Last month I spent roughly $2.600 in equivalent API usage. There is no way am paying that. I let my $100 sub lapse if next month I'll be using it less.
Look, I am not saying that there isn't a potential value out there. But the cost has to be bounded. If your opportunity is $1.000 and AI costs $2.000 to execute it, then you don't have a business model here.
You can assume Openrouter open-model providers serve at or above margin, because there's no branding so there's no reason to do it unless you can be profitable. If the Anthropic models are anywhere in that ballpark, they're very comfortably profitable on API.
That is a big exaggeration. You can have a perfectly usable local LLM setup that will power your agent for single digit thousands of dollars. Can even power multiple agents simultaneously, depending on the hardware and setup. Won't be fast and won't be frontier intelligence, but definitely useful.
Is the fraction of the 5h quote consumed consistent with the fraction of the weekly quota consumed?
I heard there is a usage tracking tool you can install that tells you if tokens are more or less expensive at the current time.
For example, if your weights were trained as 32-bit floats and you need 1TB of RAM, you could reduce that to around 256GB by quantizing to 8-bit floats. You also make the model faster in the process because there is less data to process to calculate the next token.
The game is to balance between the savings of quantization and making the model dumb enough the people notice
https://www.fool.com/investing/2026/07/25/spacexs-performanc...
Of course it's almost entirely unsubstantiated BS.
So this is a case of extraordinary claims requiring extraordinary evidence.
And even though it's super straight forward to collect the evidence, there seems to be zero substantiating the vibe bro conspiracy theories.
Another issue is also that the risk here is probably literally zero. Any evidence in support of such could easily be dismissed, with completely plausible deniability, as a short-lived technical glitch as opposed to intentional behavior.
I'm sorry, but yeah. The official account is a harness regression and some platform bugs.
Where is the evidence they are underhandedly and unethically regressing their models to shed load and reduce costs? This is the conspiracy theory running rampant through the vibe boroughs; that they are bait-and-switching on model capabilities then "nerfing" them to save money and shed load. Where is the evidence?!
I'm not saying it's illegal, per say, so don't come at me with that straw man bull cock. This bro science conspiracy has been circulating for at least 2 years(I don't even know) and enterprises would certainly be pissed off if they were paying premium API prices for advertised and previously tested model capabilities that are suddenly under performing due to "nerfing" shenanigans.
So where is the evidence?!
Not too mention these companies could easily offer one product to enteprises and another to everyone else
Model nerfing is real
I was so certain it wasn't, based on the complete lack of evidence.
But then you said it's real. NVM, I don't need evidence! Somebody said it's real!
This place has fallen off.
Bad evidence is worse than no evidence.
And you failed to address the specific criticism I made to your point, instead going for an ad hominem / poisoning the well.
In sum, you're out of line, woefully misled, and lacking in logical thinking. Three strikes, you're out.
The claim is that they nerf subscription accounts not API.
Until shit like this: https://www.anthropic.com/engineering/april-23-postmortem
Where people pointed out issues early and en masse, and Anthropic denied it was happening, gaslighted anyone claiming this was an issue, then begrudgingly admitted it was an issue, and then spent another two weeks "fixing it".
Or shit like this: https://www.anthropic.com/engineering/a-postmortem-of-three-...
Anthropic is in a perpetual state of "oops, these 'bugs' degraded our model quality" and only admit the issues when it's immediately obvious and visibly affects a large number of customers.
Otherwise all open benchmarks can be (and are) gamed. And it's quite hard to judge the output of a non-determenistic black box that Anthropic (or OpenAI) constantly tweak.
It's like arguing that your bank is scalping you by rounding down interest math on odd days of the month when they can just introduce a perfectly legal bullshit fee or otherwise change their terms to your disadvantage instead.
Banks have far greater transparency and legal requirements. LLM companies are just delivering a black box that they have complete control over. And given the regular 'How's Claude doing this session?' stuff, it's almost certain that they're A-B testing various tweaks on a per session basis.
this one has been around for over a year
I also wonder if cache could be used to throw these off as well, where it's serving un-nerfed cache results for context windows that are identical to ones they've previously had for benchmark requests.
Seems like the only way to do it well would be to have some randomness involved that couldn't be cheated on - but you'd want to do it in a way that doesn't throw out the benchmarks too much, so your results can be compared still.
All that energy wasted... could just drive a big V8 instead and make less money for the big tech
I am quite convinced that the whole nerfing phenomenon is 90% AI psychosis. I have the word muted on X.
this bench was just released, it couldn't have detected opus 4.6 degradation.
Remember that some people get access to the next flagship version long before us peasants do. I recall seeing the mention of "Astra" more than a month before it was officially announced
I would also assume they use nebulous labels like "Medium Effort" or "High Effort" map to quantitative amounts of compute allocation... and that these amounts can be varied manually or automatically. Right?
I mean, there's a reason why they call it "High Effort" and not "Exactly 5 Minutes of GPU Time on Exactly 10 GPUs." They want to be able to move those sliders and tweak those knobs.
And, while you might be billed fewer tokens as a result (because the lower thinking would result in less investigatory work), you might not know this is happening, and know to dial up effort accordingly - you'd simply get a worse work product. And certainly, Anthropic's incentive for anyone on a subscription is to push this as aggressively as they can, so people use less of that subscription.
Sadly, I'd also expect that the OP's benchmark will be detected as a test of model capabilities, and thus be given a high classification so that this strategy remains undetected.
fuck
my fair lady gained a bit weight and the bjs lack variety? [ I'm certain that that's a quote from some time ago by some commenter in a thread with a similar or even the same context ...but my Amnesia (Ｔ▽Ｔ ):･ﾟ:･ﾟ] fuck off.
you open two files, before and after you notice a nerf, and from worse comments to logical oversights, it's all damn obvious.
don't normalize this make believe bullshit and misleading people who you think barely understand what they see anyway ...
you wouldn't even know if models had somehow timed nerfs hardcoded into them, however much control over the stack you have.
ridiculous